A new paper from Meta shows that coding agents improve faster when they reuse structured summaries of past attempts instead of full execution logs. The approach reduces noise, prevents repeated failures, and cuts costs by lowering compute requirements.
Two-line summaries can outperform thousands of tokens of raw logs, addressing a key bottleneck in self-improving agent systems like Meta's HyperAgents and Meta-Harness. The philosophy shifts from scaling retries to memory compression and experience reuse.
For investors, this signals a path to better AI performance without linearly increasing costs. However, real-world results may vary as summary quality remains critical.